Hybrid KAN-MLP Architecture Improves Human Activity Recognition
A recent study introduces a hybrid framework that merges Kolmogorov-Arnold Networks (KANs) with multi-layer perceptrons (MLPs) to enhance human activity recognition (HAR) based on IMU data. While KANs perform well with clean, low-dimensional inputs, they face challenges with noisy real-world data. Conversely, MLPs exhibit greater resilience to noise and are more computationally efficient. Substituting all MLP elements with KANs in HAR models results in diminished accuracy and efficiency. The suggested architecture incorporates a KAN-based input embedding layer, maintains MLP layers for mixing intermediate features, and features a dedicated LarctanKAN module for the final classification. Tests on eight public HAR datasets validate the success of this hybrid model.
Key facts
- KANs perform well on clean, low-dimensional data but struggle with noisy real-world datasets.
- MLPs are more tolerant to noise and computationally efficient than KANs.
- Replacing all MLP components with KANs in HAR models degrades accuracy and computation efficiency.
- The proposed hybrid architecture uses a KAN-based input embedding layer.
- MLP layers are retained for intermediate feature mixing.
- A specialized LarctanKAN module is introduced for final activity classification.
- The study evaluates the architecture on eight public HAR datasets.
- The research addresses the challenge of combining KANs' precision with MLPs' noise robustness.
Entities
—